-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_torch.py
230 lines (204 loc) · 9.07 KB
/
train_torch.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
"""
https://pytorch.org/tutorials/intermediate/dist_tuto.html
"""
import torch
import torch.nn as nn
import torch.multiprocessing as mp
from torch.nn.parallel import DistributedDataParallel as DDP
import torch.distributed as dist
import numpy as np
import os
import time
import argparse
from argparse import Namespace
from tqdm import tqdm
from circuit.qnn_torch_pure import QNN
def get_opt():
parser = argparse.ArgumentParser()
parser.add_argument("--K", type=int, default=1)
parser.add_argument("--W", type=int, default=1)
parser.add_argument("--p", type=float, default=0)
parser.add_argument("--M", type=int, default=100)
parser.add_argument("--port", type=int, default=0)
parser.add_argument("--seed", type=int, default=0)
opt = parser.parse_args()
return opt
""" Gradient averaging. """
def average_gradients(model):
size = float(dist.get_world_size())
for param in model.parameters():
dist.all_reduce(param.grad.data, op=dist.ReduceOp.SUM)
param.grad.data /= size
""" Weight averaging. """
def average_weights(model):
size = float(dist.get_world_size())
for param in model.parameters():
dist.all_reduce(param.data, op=dist.ReduceOp.SUM)
param.data /= size
def step(model, feat, label, optimizer, K):
local_batch_size = len(feat) // K
for i in range(0, len(feat), local_batch_size):
loss = 0
for j in range(local_batch_size):
loss += (model(feat[i+j]) - label[i+j])**2
optimizer.zero_grad()
loss.backward()
optimizer.step()
def evaluate(model, feat_test, label_test):
n_correct = 0
feat_test = torch.from_numpy(feat_test.astype(np.complex64))
for i in range(len(feat_test)):
predict = model(feat_test[i])
if (predict.item()-0.5) * label_test[i] > 0:
n_correct += 1
return n_correct / len(feat_test)
def train(model, feat, label, feat_test=None, label_test=None, batch_size=4, local_iter=8, p=0.1, M=100, seed=0):
model.train()
acc = evaluate(model, feat_test, label_test)
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD([{'params': model.parameters(), 'lr': 1e-2, 'weight_decay': 0}], momentum=0.9)
#optimizer = torch.optim.Adam([{'params': model.parameters(), 'lr': 1e-2, 'weight_decay': 0}])
#optimizer = torch.optim.RMSprop([{'params': model.parameters(), 'lr': 1e-2, 'weight_decay': 0}])
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 40, gamma=0.5, last_epoch=-1)
loss_list = []
grad_list = []
acc_list = []
n_iter = 0
for i in range(100):
index = np.random.permutation(len(feat))
feat, label = feat[index], label[index]
loss = 0
for j in range(0, len(feat), batch_size):
feat_batch = torch.from_numpy(feat[j:j+batch_size].astype(np.complex64))
label_batch = torch.from_numpy(label[j:j+batch_size]).long().to(model.device)
loss = 0
#for m in range(local_iter):
# loss = 0
for k in range(len(feat_batch)):
predict = model(feat_batch[k])
predict = (predict - 0.5)*2
loss += (label_batch[k] - predict)**2
loss = loss / len(feat_batch)
optimizer.zero_grad()
loss.backward()
grad_list.append(np.linalg.norm([param.grad.numpy() for param in model.parameters()][0].flatten()))
optimizer.step()
#local_batch_size = len(feat_batch) // local_iter + (len(feat_batch) % local_iter != 0)
#for k in range(0, len(feat_batch), local_batch_size):
# loss = 0
# for m in range(k, min(k+local_batch_size, len(feat_batch))):
# predict = model(feat_batch[m])
# predict = (predict - 0.5)*2
# loss += (label_batch[m] - predict)**2
# #cnt += 1
# loss = loss / (min(k+local_batch_size, len(feat_batch)) - k)
# optimizer.zero_grad()
# loss.backward()
# grad_list.append(np.linalg.norm([param.grad.numpy() for param in model.parameters()][0].flatten()))
# optimizer.step()
#if dist.get_rank() == 0:
# print('Batch: {}, size: {}'.format(time.time() - st, cnt))
#average_gradients(model)
n_iter += 1
# if n_iter%local_iter == 0:
# average_weights(model)
# if dist.get_rank() == 0:
acc = evaluate(model, feat_test, label_test)
acc_list.append(acc)
scheduler.step()
loss_list.append(loss.item())
# if dist.get_rank() == 0:
# print('Epoch: {}, loss = {}'.format(i, loss))
# torch.save(model.cpu().state_dict(), 'logs/model.pth')
if torch.cuda.device_count() > 0:
model.cuda()
log_dir = 'logs/qnn/'
if not os.path.exists(log_dir):
os.makedirs(log_dir)
average_weights(model)
if dist.get_rank() == 0:
acc = evaluate(model, feat_test, label_test)
acc_list.append(acc)
#with open('logs/basis00/e200/acc_test_'+str(dist.get_world_size())+'_'+str(local_iter)+'_'+str(p)+'_'+str(M)+'.txt', 'w') as f:
# f.write(str(acc)+'\n')
np.save(log_dir+'acc_test_'+str(dist.get_world_size())+'_'+str(local_iter)+'_'+str(p)+'_'+str(M)+'_'+str(seed), acc_list)
acc_train = evaluate(model, feat, label)
with open(log_dir+'acc_train_'+str(dist.get_world_size())+'_'+str(local_iter)+'_'+str(p)+'_'+str(M)+'_'+str(dist.get_rank())+'_'+str(seed)+'.txt', 'w') as f:
f.write(str(acc_train)+'\n')
np.save(log_dir+'loss_'+str(dist.get_world_size())+'_'+str(local_iter)+'_'+str(p)+'_'+str(M)+'_'+str(dist.get_rank())+'_'+str(seed), loss_list)
np.save(log_dir+'grad_'+str(dist.get_world_size())+'_'+str(local_iter)+'_'+str(p)+'_'+str(M)+'_'+str(dist.get_rank())+'_'+str(seed), grad_list)
def parallel_train(rank, world_size, K, p, M, seed=0):
# load data
#feat_file = 'data/wine.data'
#feat, label = [], []
#with open(feat_file, 'r') as f:
# for line in f.readlines():
# parts = line.strip().split(',')
# if float(parts[0]) > 2:
# continue
# label.append((float(parts[0])-1)*2-1)
# feat.append([float(part) for part in parts[1:]])
#data_len = len(feat) // world_size + 1
#feat = np.array(feat, dtype=np.float32)[rank*data_len:(rank+1)*data_len, np.newaxis, :]
#label = np.array(label)[rank*data_len:(rank+1)*data_len]
#data = np.loadtxt('data/iris_classes1and2_scaled.txt')
#index = np.random.permutation(len(data))
#data = data[index]
#data_len = len(data) // world_size + (len(data) % world_size != 0)
#label = data[:, -1][rank*data_len:(rank+1)*data_len]
#feat = data[rank*data_len:(rank+1)*data_len, np.newaxis, :-1]
#feat = np.concatenate((feat, feat), axis=-1)
#print(feat.shape)
feat_train = np.load('data/mnist_train_feat.npy')
#data_len = len(feat_train) // world_size + (len(feat_train) % world_size != 0)
data_len_min = len(feat_train) // world_size
offset = len(feat_train) % world_size
if rank < offset:
start = rank*(data_len_min+1)
end = start+data_len_min+1
else:
start = offset*(data_len_min+1)+(rank-offset)*data_len_min
end = start+data_len_min
feat_train = feat_train[start:end]
label_train = np.load('data/mnist_train_label.npy')
label_train = label_train[start:end]
feat_test = np.load('data/mnist_test_feat.npy')
label_test = np.load('data/mnist_test_label.npy')
# define model
n_qubits = 6 #feat_train.shape[-1]
n_layers = 4
torch.manual_seed(seed)
np.random.seed(seed)
model = QNN(n_qubits, n_layers, p=p, M=M, param_shift=True)
# average_weights(model)
# parallel wrapper
if torch.cuda.device_count() > 0:
model = nn.DataParallel(model)
model.cuda()
#batch_size = 64 // world_size
batch_size = 64
train(model, feat_train, label_train, feat_test, label_test, batch_size=batch_size, local_iter=K, p=p, M=M, seed=seed)
def init_process(rank, size, K, port, p, M, seed, fn, backend='gloo'):
""" Initialize the distributed environment. """
os.environ['MASTER_ADDR'] = '127.0.0.1'
os.environ['MASTER_PORT'] = str(port)
dist.init_process_group(backend, rank=rank, world_size=size)
torch.set_num_threads(1)
fn(rank, size, K, p, M, seed=seed)
if __name__ == '__main__':
opt = get_opt()
# mp.set_start_method("spawn")
# cost = []
# for world_size in tqdm([opt.W], desc='world_size'):
# #world_size = 2#mp.cpu_count()
# processes = []
# st = time.time()
# for rank in range(world_size):
# p = mp.Process(target=init_process, args=(rank, world_size, opt.K, opt.port+29500, opt.p, opt.M, opt.seed, parallel_train))
# p.start()
# processes.append(p)
# for p in processes:
# p.join()
# cost.append(time.time() - st)
# np.save('logs/qnn/time'+str(opt.W)+'_'+str(opt.K)+'_'+str(opt.p)+'_'+str(opt.M)+'_'+str(opt.seed), cost)
parallel_train(0, 1, 1, 0, 0, seed=0)